Intelligent flood monitoring framework using dual-polarization synthetic aperture radar and geospatial artificial intelligence

ABSTRACT Flood maps are needed quickly during monsoon disasters, when cloud blocks optical satellite images. Agencies in South Asia still lack a clear comparison of established radar methods and of the two Sentinel-1 polarizations, for agricultural floodplains. This study compares four methods on the Kovvur reach of the Lower Godavari River Basin, India, for the August 2019 flood. Texture analysis, dual-layer backscatter thresholding, image differencing and an artificial neural network were applied to the same dual-polarized Sentinel-1 images. Accuracy was recalculated from the 500-pixel confusion counts, using National Remote Sensing Centre maps as the reference. On those counts, the neural network had the highest overall accuracy, 89.40%, and a flood area of 77.09 km2, close to the reference area of 79.50 km2. Thresholding was lower, at 82.00%. These scores are descriptive. The locations are paired, the reference source is shared with training, and a paired significance test could not be computed. The contribution is a practical comparison for this Lower Godavari agricultural floodplain, limited to one flood and one reach.